Abstract
This paper presents a fast two-stage corner detector with noise tolerance. In the first stage, a novel candidate pruning approach based on PSO-SVM is proposed to select candidatecorner pixels which have great potential to be corners. In the second stage the Harris corner detector is employed to recognize real corners among the candidate-corner pixels. The parameters and feature selection of SVM classifier is optimized by using particle swarm optimization (PSO). The method takes advantage of the minimum structure risk of SVM and the quickly globally optimizing ability of PSO. Generally speaking, corners are considered as the junction of edges. Thus, edge pixels with a high gradient in more than one direction should be selected as candidate corners. Meanwhile, impulse noise often corrupts digital images while images are transmitted over an unreliable channel or are captured using a camera with faulty sensors. Noise-corrupted pixels usually cause serious false detection problems in most corner detectors. The proposed PSO-SVM candidate pruning approach detects noisy pixels and excludes them from being candidate corners to enhance the noise tolerance of the corner detector. Through the well-selection of candidate corners, the proposed candidate pruning approach can 1) enhance the noise tolerance capability, and 2) reduce the computational effort of the corner detectors.